Abstract: To address the problems of stage-dependent degradation, strong noise interference in oil debris signals, and difficulty in continuously characterizing gear health states over the full life cycle, a gear wear health index construction method is proposed by integrating Unscented Kalman Filter (UKF)-based temporal preprocessing and a sliding-window Variational Autoencoder (VAE). First, oil ferrographic images are used to extract features including IPCA, relative debris concentration, total particle count, and the number of particles larger than 80 μm, which jointly describe wear coverage, concentration level, and severe particle release characteristics. Second, a baseline-guided UKF with forward filtering and backward smoothing is introduced to suppress isolated spikes while preserving the intrinsic degradation trend of multivariate debris signals. Based on the filtered features, a cumulative wear metric is constructed to capture the long-term degradation history of the gear system. Subsequently, a sliding-window strategy is adopted to form sequential samples, and a VAE is trained using data from the normal wear stage. The reconstruction error is used to quantify deviations from normal operating patterns, and is further combined with cumulative wear constraints to generate a continuous health index. Experimental results demonstrate that the proposed method effectively characterizes the full-life degradation process of gear wear from run-in to severe wear, while improving the stability and robustness of health assessment